From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detection

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "From Text Metrics to Model Internals".

Jane: Here's the gist of the paper. The gist: Probing Whisper’s intermediate decoder states yields highly competitive detection via a BLSTM (F1 score of 62.1%) without reference transcripts.

Tom: Text-Based Detection The study investigates the discriminative power of text-based metrics, categorizing them into oracle and reference-free methods The most commonly utilized metrics require a ground-truth transcript to measure deviations in the ASR hypothesis These include traditional error metrics such as Word Error Rate (WER), Character Error Rate (CER), Insertion Error Rate (IER)

22: , and the reference vs.

Jane: prediction Length Ratio (LenR)

23: To capture deeper semantic divergence, BERTScore (BERT)

10: and SeMaScore (SeMa)

9: are also employed Reference-free metrics are less commonly used but are particularly important, as operating solely on the ASR prediction and audio metadata makes them applicable in zero-shot deployments These include the 5-gram Repetition rate (Rep)

8: , Stopword Ratio (StopW), and GPT 2 Perplexity (PPL)

14: The analysis showed that tree-based ensemble classifiers substantially outperform linear baselines, and that reference-free text features alone remain insufficient for reliable detection Specifically, without semantic and structural reference anchors, the models became overly conservative, suffering a collapse in Recall (Logistic Regression Recall: 15.5%) LLM-Based Detection The research evaluates Large Language Models (LLMs) for zero-shot ASR error correction by prompting them to strictly isolate semantic fabrications from acoustic mishearings Out-of-the-box LLMs struggle significantly, with both achieving an F1 score of roughly 41% However, injecting model-specific pathology knowledge into the prompt substantially improves precision Combining this failure taxonomy with domainspecific few-shot examples yielded the highest overall LLM performance (F1: 58.7%, Accuracy: 88.4%) Despite extensive prompt engineering and in-context learning, LLMs failed to surpass the XGBoost text-feature classifier (F1: 62.8%) Furthermore, when transitioning to the referencefree setting, LLM performance suffered a collapse, dropping to an F1 score of 32.8% This mirrors the severe degradation observed in the reference-free text classifier, demonstrating that despite the extensive semantic knowledge embedded in LLMs, zero-shot hallucination detection remains fundamentally constrained by the absence of ground-truth anchors Internal State Probing The investigation into Whisper’s decoder internal states revealed that hallucination signals are encoded across intermediate layers of the Whisper decoder and that probing full decoding sequences consistently outperforms single-token embeddings generated at the end of the decoding process Early layers (0-5) as well as late ones (27-31) perform significantly worse in separability, while middle layers show little variance To capture these sequential dependencies, a Bidirectional Long Short-Term Memory (BLSTM) classifier was utilized on the entire decoding sequence embeddings

29: The optimal BLSTM configuration achieved an AUC of 87.6% (F1: 65.5%, Precision: 68.1%, Recall: 63.4%) This reference-free internal state detector outperforms the best text-feature models (F1: 62.8%) despite not relying on ground-truth transcripts Detector Fusion Analysis of detection overlap revealed that the paradigms capture non-overlapping signals (agreement: 0.64-0.73) The text classifier maximizes Recall (0.73) at the cost of Precision (0.53), while the LLM is the most conservative, with the highest Precision (0.64) but the lowest Recall (0.51) The internal state detector achieves the highest F1 score (0.62), being the most balanced A late-fusion stacking ensemble combining the outputs of both detectors into a lightweight meta-classifier achieved the highest overall performance for hallucination detection on this dataset (F1 score of 68.3%) This approach establishes the highest overall performance for hallucination detection on this dataset, highlighting the potential of cross-paradigm detection Robust detection of ASR hallucinations remains challenging, particularly in zero-shot deployments where ground-truth references are unavailable Our investigation across three paradigms revealed that while text-based classifiers and LLMs achieve strong oracle performance, both suffer performance collapses in strictly reference-free settings We demonstrate that probing Whisper’s intermediate decoder states yields highly competitive detection via a BLSTM (F1 score of 62.1%) without reference transcripts While a late-fusion meta-classifier combining text and internal state probabilities achieves our highest overall performance (F1 score of 68.3%), it forfeits zeroshot usability Ultimately, internal state probing, despite requiring direct model access, provides a highly promising solution Our internal-state analysis is limited to Whisper, but this approach applies to other autoregressive encoder-decoder ASR models and, more broadly, to any architecture that exposes hidden states during token-by-token decoding Extending internal-state probing to alternative ASR models and paradigms remains an important direction for future work Furthermore, utilizing a specialized, lightweight LLM fine-tuned specifically for hallucination detection would bridge the gap between the zero-shot commercial LLMs tested and our supervised text-feature classifiers Acknowledgments This research was supported by the National Science Centre, Poland under Grants 2021/42/E/ST7/00452 and 2023/49/B/ST7/04100, and by program ”Excellence initiative – research university” for the AGH University of Krakow We gratefully acknowledge Polish high-performance computing infrastructure PLGrid (HPC Centers: ACK Cyfronet AGH) for providing computer facilities and support within computational Grants PLG/2025/018238 and PLG/2025/018153 Generative AI Use Disclosure The authors used large language models (ChatGPT, Gemini, Claude) to assist with language editing All experimental design, data processing, statistical analysis, and scientific conclusions were independently conducted and verified by the authors The authors take full responsibility for the content of this manuscript References

1: A.

Tom: Radford, J.

Title and authors: Tom: We’re moving into the actual summary of "From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detection," which outlines the specific setup they used for this investigation.

Jane: So, they started by using a human-annotated dataset called HALAS, which is crucial because it gives them those specific hallucination labels for predictions from seven state-of-the-art models using Earnings audio data.

Lu: They focused on the Whisper large v3 model specifically, looking at eight hundred fifty-eight out of three thousand six hundred eleven predictions marked as hallucinations, which is about twenty-three point eight percent of the total predictions in that set <ref:2606.23060#pg1>.

Meng: And they also looked at other error types, like looping phrases—which they found were only about zero point five percent—and the rest were just clean predictions without hallucinations present in that specific way <ref:2606.23060#pg1>.

Tom: They then set up their three detection paradigms: text-based metrics, LLM-based detection with prompt conditioning, and finally the internal decoder state probing using a Bidirectional Long Short-Term Memory classifier on the entire decoding sequence embeddings.

Jane: The key finding they present here is that while text classifiers score high on recall when they have reference transcripts, those same text features degrade significantly without any ground-truth reference available for comparison.

Lu: And the LLM results showed that injecting domain-specific pathology knowledge into the prompt substantially improved precision, reaching an F1 score of fifty-eight point seven percent when they combined that failure taxonomy with specific few-shot examples from the HALAS training split <ref:2606.23060#pg1>.

Meng: But even with those enhancements, the LLM still didn't manage to surpass the best text-feature model in terms of overall F1 score, which was sixty-two point eight percent <ref:2606.23060#pg1>.

Tom: So, if you take it all together, the summary is that they benchmarked three distinct detection strategies against a specific dataset to see where each one actually performs best when dealing with ASR hallucinations.

Jane: And the big implication of this summary is clearly pointing toward internal model representations as a more reliable source for error detection than just relying on the text output or external knowledge alone.

The paper's summary: Tom: Now, let's look at what improvements the authors suggest based on their findings in "From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detection."

Jane: They suggest that combining different detection paradigms is actually beneficial because they found that the three methods capture non-overlapping signals, with agreement rates between zero point six four and zero point seven three.

Lu: The text classifier excels at recall, achieving a recall of zero point seven three, but it comes at the cost of precision, settling around zero point five three when reference-free conditions are applied to that method specifically.

Meng: The LLM approach is the most conservative in terms of precision, hitting a precision score of zero point six four, but it sacrifices recall down to zero point five one because it’s too cautious without those ground truth anchors.

Tom: And the paper points out that the internal state detector achieves the highest F1 score at zero point six two and is described as the most balanced approach across these trade-offs between precision and recall.

Jane: So, they are suggesting that instead of picking just one method, using a late fusion stacking ensemble combining both text and internal state probabilities into a meta-classifier can achieve the highest overall performance on this dataset.

Lu: That late fusion meta-classifier achieved an F1 score of sixty-eight point three percent overall, which is their highest result on this specific test split, demonstrating that combining these different detection signals is more effective than any single method alone <ref:2606.23060#pg1>.

Meng: From a practical standpoint, though they admit that this top performance comes at the cost of zeroshot usability because it relies on having both the text and internal state outputs available.

Tom: So, the suggested improvement isn't just one better metric; it’s a combination strategy where you leverage the strengths of each paradigm to get that sixty-eight point three percent score <ref:2606.23060#pg1>.

Jane: It really highlights how complementary these different detection methods are, even though they seem to focus on different aspects of the problem—text versus model internals.

The paper's improvements: Tom: We’re coming to the conclusion of "From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detection" and summarizing what this all means for the field.

Jane: Basically, while text-based classifiers and LLMs show strong oracle performance when they have ground truth, both methods suffer major performance drops when you move into strictly reference-free settings.

Lu: The key finding that stands out is that probing Whisper’s intermediate decoder states using a Bidirectional Long Short-Term Memory classifier yields highly competitive detection via an F1 score of sixty-two point one percent without needing any reference transcripts whatsoever <ref:2606.23060#pg1>.

Meng: So, the main implication for deployment is that internal state probing provides a promising solution because it works without needing ground-truth transcripts to function as a detector itself, even though you need access to the model's internals.

Tom: And they also point out that while the late fusion meta-classifier gets the highest overall F1 score of sixty-eight point three percent, that approach sacrifices zeroshot usability because it requires both text and internal state information <ref:2606.23060#pg1>.

Jane: Ultimately, this paper suggests that internal state probing is a very promising path forward, even if it’s limited to Whisper right now, because it offers a detection mechanism independent of the ground-truth transcript.

Lu: It opens up a direction where this kind of probing technique could be applied to other autoregressive encoder-decoder ASR models and more broadly, any architecture that exposes its hidden states during token-by-token decoding.

Meng: And they also suggest that for the future, we might see specialized, lightweight LLMs fine-tuned specifically for hallucination detection bridging the gap between those general zero-shot LLMs and supervised text classifiers.

Tom: So, to wrap up on "From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detection," we have established that internal state probing is a highly promising way to detect hallucinations without reference transcripts, even if it means accepting some limitations on usability or requiring direct model access.

Jane: It’s a big step because it shows that the internal workings of the model can provide valuable detection signals on their own.

Lu: It definitely sets a new direction for research into how we use these hidden states to understand and correct generative speech errors.

Tom: So that’s everything we have for this paper today. We’ve looked at the text metrics, the LLMs, and finally the internal state probing method in "From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detection."

Conclusion: Tom: So we’ve looked at how they used text metrics, LLMs, and internal states to detect hallucinations in Whisper ASR using their paper "From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detection."

Jane: Right, so the main thing is that while text-based methods and LLMs are good when you have reference transcripts they struggle significantly when you take them out, which is a big issue for zero-shot systems.

Lu: The paper shows that probing Whisper’s intermediate decoder states using a Bidirectional Long Short-Term Memory classifier gives them an F1 score of sixty-two point one percent without any reference transcripts, which is pretty competitive against the best text features at sixty-two point eight percent.

Meng: But they did flag a limitation, saying that while that internal state method is good on its own, combining it with the text or LLM outputs into a late fusion ensemble gets them to an F1 score of sixty-eight point three percent, but then you lose the zero-shot capability of just using those raw outputs.

Lalam: From my perspective, the internal state probing result is interesting because it works without any external anchors, which is exactly what we need for real-world deployments where reference transcripts aren't available. I'm confident in this finding; it shows a fundamental signal within the model itself.

Tom: So they are suggesting that even though internal state probing requires direct access to the model’s decoding sequence embeddings, it offers a highly promising solution for hallucination detection on Whisper.

Jane: That makes sense. It moves the focus from just looking at what's written to actually seeing how the AI is generating the speech token by token, and that's a very different kind of signal.

Lu: They also point out that this internal state detector performs worse in early and late layers, but it’s still better than nothing if you can’t get a reference transcript.

Meng: For me at the startup level, the cost to run this internal state probe is high because you need to actually run the full decoding sequence embeddings through a BLSTM classifier every time, but if it works on Whisper, it might work on other autoregressive models too.

Lalam: I agree with Meng; that's exactly what I was thinking. It shows that we should look at architectures like this and see if those hidden states provide a more robust way to identify when the model is hallucinating.

Tom: So, "From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detection" ends with the conclusion that while internal state probing is promising without reference transcripts, combining it with other methods gives the highest overall score, even if you sacrifice some convenience.

Jane: It’s a really important piece of work because it shows that we don't have to rely solely on external ground truth to catch errors in these complex generative speech models.

Lu: The future direction they suggest is extending this internal-state probing technique to other ASR models, which is a big step for representation learning research.

Meng: I think the next challenge for us will be figuring out how much compute that BLSTM classifier needs to run efficiently at scale before we can actually deploy it in production.

Lalam: I’m ready for the next paper, though, because this whole idea of using internal states to detect errors has huge implications across multimodal AI.